303 32 6MB
German Pages [268] Year 2011
Roger Lee (Ed.) Computer and Information Science 2011
Studies in Computational Intelligence, Volume 364 Editor-in-Chief Prof. Janusz Kacprzyk Systems Research Institute Polish Academy of Sciences ul. Newelska 6 01-447 Warsaw Poland E-mail: kacprzyk@ibspan.waw.pl Further volumes of this series can be found on our homepage: springer.com Vol. 345. Shi Yu, Léon-Charles Tranchevent, Bart De Moor, and Yves Moreau Kernel-based Data Fusion for Machine Learning, 2011 ISBN 978-3-642-19405-4 Vol. 346. Weisi Lin, Dacheng Tao, Janusz Kacprzyk, Zhu Li, Ebroul Izquierdo, and Haohong Wang (Eds.) Multimedia Analysis, Processing and Communications, 2011 ISBN 978-3-642-19550-1 Vol. 347. Sven Helmer, Alexandra Poulovassilis, and Fatos Xhafa Reasoning in Event-Based Distributed Systems, 2011 ISBN 978-3-642-19723-9 Vol. 348. Beniamino Murgante, Giuseppe Borruso, and Alessandra Lapucci (Eds.) Geocomputation, Sustainability and Environmental Planning, 2011 ISBN 978-3-642-19732-1 Vol. 349. Vitor R. Carvalho Modeling Intention in Email, 2011 ISBN 978-3-642-19955-4 Vol. 350. Thanasis Daradoumis, Santi Caball´e, Angel A. Juan, and Fatos Xhafa (Eds.) Technology-Enhanced Systems and Tools for Collaborative Learning Scaffolding, 2011 ISBN 978-3-642-19813-7 Vol. 351. Ngoc Thanh Nguyen, Bogdan Trawi´nski, and Jason J. Jung (Eds.) New Challenges for Intelligent Information and Database Systems, 2011 ISBN 978-3-642-19952-3 Vol. 352. Nik Bessis and Fatos Xhafa (Eds.) Next Generation Data Technologies for Collective Computational Intelligence, 2011 ISBN 978-3-642-20343-5 Vol. 353. Igor Aizenberg Complex-Valued Neural Networks with Multi-Valued Neurons, 2011 ISBN 978-3-642-20352-7 Vol. 354. Ljupco Kocarev and Shiguo Lian (Eds.) Chaos-Based Cryptography, 2011 ISBN 978-3-642-20541-5
Vol. 355. Yan Meng and Yaochu Jin (Eds.) Bio-Inspired Self-Organizing Robotic Systems, 2011 ISBN 978-3-642-20759-4 Vol. 356. Slawomir Koziel and Xin-She Yang (Eds.) Computational Optimization, Methods and Algorithms, 2011 ISBN 978-3-642-20858-4 Vol. 357. Nadia Nedjah, Leandro Santos Coelho, Viviana Cocco Mariani, and Luiza de Macedo Mourelle (Eds.) Innovative Computing Methods and their Applications to Engineering Problems, 2011 ISBN 978-3-642-20957-4 l Vol. 358. Norbert Jankowski, Wlodzis aw Duch, and Krzysztof Grabczewski (Eds.) ֒ Meta-Learning in Computational Intelligence, 2011 ISBN 978-3-642-20979-6 Vol. 359. Xin-She Yang, and Slawomir Koziel (Eds.) Computational Optimization and Applications in Engineering and Industry, 2011 ISBN 978-3-642-20985-7 Vol. 360. Mikhail Moshkov and Beata Zielosko Combinatorial Machine Learning, 2011 ISBN 978-3-642-20994-9 Vol. 361. Vincenzo Pallotta, Alessandro Soro, and Eloisa Vargiu (Eds.) Advances in Distributed Agent-Based Retrieval Tools, 2011 ISBN 978-3-642-21383-0 Vol. 362. Pascal Bouvry, Horacio González-Vélez, and Joanna Kolodziej (Eds.) Intelligent Decision Systems in Large-Scale Distributed Environments, 2011 ISBN 978-3-642-21270-3 Vol. 363. Kishan G. Mehrotra, Chilukuri Mohan, Jae C. Oh, Pramod K. Varshney, and Moonis Ali (Eds.) Developing Concepts in Applied Intelligence, 2011 ISBN 978-3-642-21331-1 Vol. 364. Roger Lee (Ed.) Computer and Information Science, 2011 ISBN 978-3-642-21377-9
Roger Lee (Ed.)
Computer and Information Science 2011
123
Editor
Guest Editors
Prof. Roger Lee
Wencai Du Simon Xu
Central Michigan University Computer Science Department Software Engineering & Information Technology Institute Mt. Pleasant, MI 48859 U.S.A. E-mail: lee1ry@cmich.edu
ISBN 978-3-642-21377-9
e-ISBN 978-3-642-21378-6
DOI 10.1007/978-3-642-21378-6 Studies in Computational Intelligence
ISSN 1860-949X
c 2011 Springer-Verlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typeset & Cover Design: Scientific Publishing Services Pvt. Ltd., Chennai, India. Printed on acid-free paper 987654321 springer.com
Preface
The purpose of the 10th International Conference on Computer and Information Science(ICIS 2011) held on May 16-18, 2011 Sanya, Hainan Island, China was to bring together researchers and scientists, businessmen and entrepreneurs, teachers and students to discuss the numerous fields of computer science, and to share ideas and information in a meaningful way. Our conference officers selected the best 20 papers from those papers accepted for presentation at the conference in order to publish them in this volume. The papers were chosen based on review scores submitted by members of the program committee, and underwent further rounds of rigorous review. In Chapter 1, Honghao Gao et al. In this paper, to overcome these two challenges, we primarily present an On-the-fly approach to modeling the Web navigation behaviors, and apply bounded model checking (BMC) to verifying the On-the-fly navigation model. Finally, a prototype system is discussed. In Chapter 2, Rahma Fourati et al. Within this context, we propose an approach that identifies anti-patterns in UML designs through the use of existing and newly defined quality metrics. Operating at the design level, our approach examines structural and behavioral information through the class and sequence diagrams. It is illustrated through five, well-known anti-patterns: Blob, Lava Flow, Functional Decomposition, Poltergeists, and Swiss Army Knife. In Chapter 3, Juxiang Zhou et al. In this paper a novel feature extraction approach is proposed for facial expression recognition by using the curvelet and the LDP (Local Directional Pattern). First, the low frequency coefficients of Curvelet decomposition on expression region are selected as global facial features. Then, LDP descriptor is used to describe eyes region and mouth region respectively as local facial features. In Chapter 4, Jinhui Yuan et al. In the scenario sensor nodes have the capability to adjust their transmission power with the transmission range, we approximately construct the maximum lifetime data gathering tree with the goal to balance the energy consumption among the sensor nodes to prolong the lifetime of the network. Our simulation shows that our approach is effective.
VI
Preface
In Chapter 5, Dapeng Liu et al. The experiment results demonstrate that while novice programmers are diverse in terms of programming styles, good ones tend to control execution in finer granularity. Source code format can be a flag of programming performance. It seems that there is no direct correlation between the frequency of keystrokes and the quality of programs. In Chapter 6, Kem Z.K. Zhang et al. In this study, we adopt a theory of planned behavior perspective and build up a switching model to explain blog service switching behavior. We employ a survey to explain how two quality beliefs (service quality and quality of alternatives) and two types of costs (sunk costs and relationship costs) exert influence in determining bloggers’ switching behavior. Discussions and implications are provided to better understand the switching behavior of blog and other social technologies. In Chapter 7, Kem Z.K. Zhang et al. In this study, we shed light on the user contribution of online review platforms. In particular, we attempt to understand how information technology (IT) enabled features can facilitate users’ online review contribution. To achieve this objective, we conduct a preliminary study on a Chinese online review platform. The findings confirm that social networking technology and virtual community technology provide helpful IT-enabled features to attain a high level of user contribution on the platform. Implications for both researchers and practitioners are discussed. In Chapter 8, Toukir Imam et al. This paper contains a detail description of the Fuzzy12 algorithm, its implementation and the set-up of the EDG simulation. Our simulation results demonstrate that the Fuzzy algorithm outperforms the LRU replacement algorithm with different perspective, for example, hit ratio, byte hit ratio, miss rate etc. In Chapter 9, Natalia C. Silva et al. In this paper, we propose a methodology to tackle such a problem by naturally moving from informal business rules toward the implementation of abusiness process using complex event processing. The methodology allows for theactive participation of business people at all stages of the refinement process. This is important to guarantee the correct alignment between information systems and business needs. Throughout the paper, we present an example to illustrate the application of the methodology. The methodology was applied to implement a realprocess of a building company. In Chapter 10, Hartinder Singh Johal et al. This manuscript tends to realize the above situation by proposing a three dimensional network address translation scheme with a tentative capability to support end-to-end connectivity based applications and at the same time retaining the benefits of conventional NAT model.
Preface
VII
In Chapter 11 Majid Rafigh et al. This paper proposed a new routing algorithm schema based on event occurrence pattern to satisfying k-coverage of event paths and maintaining degree of coverage in maximum level as more as possible. This method improves the network lifetime by shifting the routing responsibility from covering nodes to communication nodes, while maximizing the degree of coverage in the main path of events. In Chapter 12 Jin Zhang et al. This specification formally defines the regular behavior and fault tolerance behavior of priority queue. In particular, a priorityconcatenation operator is defined to handle the ordering of data items to ensure the highest-priority item is removed first. A finite state machine as an implementation is built based on this specification. In addition, we also discuss a priority upgrading approach to handle possible starvation situation of low-priority data items in the priority queue. In Chapter 13 Ming Ke et al. The global statistical properties of the network revealed the brain functional network for chewing of gum had small-world effect and scale-free property. Computing the degree and betweenness which belong to the centrality indices, we found that the neocortical hubs of the network were distributed in the sense and motor cortex, and the nodes in the thalamus and lentiform nucleus held the largest betweenness. The sense and motor cortices as well as thalamus and lentiform nucleus have the important roles in dispatch and transfer information of network. In Chapter 14 Yosuke Motoki et al. In this paper, we propose a new mechanism based on GSP that is used in advertisement auctions. Each advertisement has some value, because users click the advertisement when it may be useful for them. We analyze the auctioneer’s profit in comparison between normal GSP, normal VCG (Vickrey-Clarke-Groves Mechanism) and our proposed mechanism. The contribution of our research includes to clarify the features and advantages of advertisement auctions and effects to website owner’s profit rate. In Chapter 15 Li Yunpeng et al. As a theoretical creation, the research explores the coordinating pattern of value chain and has constructed the model of the value chain platform in tourism destinations and has proposed relevant theories. That is supporting the destination value chain operate effectively with dynamic integrated technology of tourism information system construction; the data searching and analyzing approaches to efficiently process the feedbacks from data users; studying the satisfaction level of tourism guests. In Chapter 16 Ge Zhu et al. The result shows that China Mobile users’ switching costs are significantly higher than those for customers of China Unicom, and the gap was increasing generally. The quantitative analysis demonstrates that reducing of consumer switching costs will relatively benefit small operators and intensify competition.
VIII
Preface
In Chapter 17 Hui Zhou et al. In our experiments, to identify a large ISP cloud, we spread vantage points inside the cloud and over the world, and collect topology information by probing a fixed list of IP addresses which consists of more than 25,000 routers and 36,000 links. Data analysis shows that sampling bias, if undetected, could significantly undermine the conclusions drawn from the inferred topologies. In Chapter 18 Hui Zhou et al. The experiment result indicates that, after applying the object snapshot concept of software, RichMap can smoothly capture and present complete router-level snapshots and significantly decrease the network load that it generates. In Chapter 19 Su Chen et al. The experimental results indicate that a sequential genetic algorithm with intensive interactions can be accelerated by being translated into CUDA code for GPU execution. In Chapter 20 Haeng-Kon Kim. In this paper, we design and Implement a sensor framework systems related to medical and surveillance that are significantly considered for enhancing human life. These are employed under USN environment to construct multiple health care services in which medical sensors are inter-connected to provide efficient management of them. It is our sincere hope that this volume provides stimulation and inspiration, and that it will be used as a foundation for works yet to come. May 2011
Guest Editors George Du Simon Xu
Contents
Applying Bounded Model Checking to Verifying Web Navigation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Honghao Gao, Huaikou Miao, Shengbo Chen, Jia Mei
1
A Metric-Based Approach for Anti-pattern Detection in UML Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rahma Fourati, Nadia Bouassida, Hanˆene Ben Abdallah
17
A Novel Feature Extraction for Facial Expression Recognition via Combining the Curvelet and LDP . . . . . . . . . . . Juxiang Zhou, Tianwei Xu, Yunqiong Wang, Lijin Gao, Rongfang Yang
35
Constructing Maximum-Lifetime Data Gathering Tree without Data Aggregation for Sensor Networks . . . . . . . . . . . . . . Jinhui Yuan, Hongwei Zhou, Hong Chen
47
An Empirical Study of Programming Performance Based on Keystroke Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dapeng Liu, Shaochun Xu
59
A Theory of Planned Behavior Perspective on Blog Service Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kem Z.K. Zhang, Sesia J. Zhao, Matthew K.O. Lee, Huaping Chen
73
User Contribution and IT-Enabled Features of Online Review Platforms: A Preliminary Study . . . . . . . . . . . . . . . . . . . . . Kem Z.K. Zhang, Sesia J. Zhao, Matthew K.O. Lee, Huaping Chen
85
X
Contents
Implementation and Performance Analysis of Fuzzy Replica Replacement Algorithm in Data Grid . . . . . . . . . . . . . . . Toukir Imam, Rashedur M. Rahman
95
Integrating Business Process Analysis and Complex Event Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Nat´ alia C. Silva, Cec´ılia L. Sabat, C´esar A.L. Oliveira, Ricardo M.F. Lima 3D NAT Scheme for Realizing Seamless End-to-End Connectivity and Addressing Multilevel Nested Network Address Translation Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 Hartinder Singh Johal, Balraj Singh, Amandeep Nagpal, Kewal Krishan Maximizing Coverage Degree Based on Event Patterns in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Majid Rafigh, Maghsoud Abbaspour Formal Specification and Implementation of Priority Queue with Starvation Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 Jin Zhang, Gongzhu Hu, Roger Lee Brain Functional Network for Chewing of Gum . . . . . . . . . . . . . . 169 Ming Ke, Hui Shen, Zongtan Zhou, Xiaolin Zhou, Dewen Hu, Xuhui Chen Effects of Value-Based Mechanism in Online Advertisement Auction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 Yosuke Motoki, Satoshi Takahashi, Yoshihito Saito, Tokuro Matsuo Research on Dynamic Optimized Approach of Value Chain in Tourist Destinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Li Yunpeng, Xie Yongqiu, Ni Min, Hao Yu, Qi Lina Analysis and Quantitative Calculation on Switching Costs: Taking 2002-2006 China Wireless Telecommunications Market as an Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 Ge Zhu, Jianhua Dai, Shan Ao An Empirical Study of Network Topology Inference . . . . . . . . . 213 Hui Zhou, Wencai Du, Shaochun Xu, Qinling Xin Computer Network Reverse Engineering . . . . . . . . . . . . . . . . . . . . . 227 Hui Zhou, Wencai Du, Shaochun Xu, Qinling Xin
Contents
XI
CUDA-Based Genetic Algorithm on Traveling Salesman Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 Su Chen, Spencer Davis, Hai Jiang, Andy Novobilski Design and Implementation of Sensor Framework for U-Healthcare Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 Haeng-Kon Kim Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
List of Contributors
Maghsoud Abbaspour Shahid Beheshti University, Iran Hanêne Ben Abdallah University of Sfax, Tunisia Email: hanene.benabdallah@fsegs.rnu.tn Nadia Bouassida University of Sfax, Tunisia Email: nadia.bouassida@isimsf.rnu.tn Shan Ao Beijing Info. Science & Tech University, China Hong Chen Renmin University of China Email: chong@ruc.edu.cn Huaping Chen University of Science and Technology of China, China Email: hpchen@ustc.edu.cn Shengbo Chen Shanghai University, China Email: schen@shu.edu.cn Su Chen Arkansas State University, U.S.A. Email: su.chen@smail.astate.edu Xuhui Chen Lanzhou University of Technology, China Email: xhchen@lut.cn Jianhua Dai Beijing Info. Science & Tech University, China
Spencer Davis Arkansas State University, U.S.A. Email: spencer@smail.astate.edu Wencai Du Hainan University, China Email: wencai@hainu.edu.cn Rahma Fourati University of Sfax, Tunisia Email: rahma.fourati10@gmail.com Honghao Gao Shanghai University, China Email: gaohonghao@shu.edu.cn Lijin Gao Yunnan Normal University, China Dewen Hu National University of Defense Technology, China Gongzhu Hu Central Michigan University, USA E-mail: hu1g@cmich.edu Toukir Imam North South University, Bangladesh Hai Jiang Arkansas State University, U.S.A. Email: hjiang@smail.astate.edu Hartinder Singh Johal Lovely Professional University, India Email: hs.johal@lpu.co.in Ming Ke Lanzhou University of Technology, China
XIV
List of Contributors
Haeng-Kon Kim Catholic University of Daegu, Korea Email: hangkon@cu.ac.kr
Andy Novobilski Arkansas State University, U.S.A. Email: anovobilski@smail.astate.edu
Kewal Krishan Lovely Professional University, India Email: kewal.krishan@lpu.co.in
C´esar A. L. Oliveira Federal University of Pernambuco, Brazil Email: calo@cin.ufpe.br
Matthew K.O. Lee City University of Hong Kong, China Email: ismatlee@cityu.edu.hk Roger Lee Central Michigan University, USA Email: lee1ry@cmich.edu Ricardo M.F. Lima Federal University of Pernambuco, Brazil Email: rmfl@cin.ufpe.br Qi Lina Beijing University, China Dapeng Liu The Brain Tech., U.S.A. Email: dliu@the brain.com Tokuro Matuso Yamagata University, Japan Jia Mei School of Computer Engineering and Science Shanghai University, China Email: me269@shu.edu.cn Huaikou Miao Shanghai University, China Email: hkmiao@shu.edu.cn Ni Min Beijing Yong You Software Company, China
Majid Rafigh Shahid Beheshti University, Iran Email: m.rafigh@mail.sbu.ac.ir Rashedur M. Rahman North South University, Bangladesh Cecília L. Sabat Federal University of Pernambuco, Brazil Email: cls@cin.ufpe.br Yoshihito Saito Yamagata University, Japan Hui Shen National University of Defense Technology Changsha, China Natália C. Silva Federal University of Pernambuco, Brazil Email: ncs@cin.ufpe.br Balraj Singh Lovely Professional University, India Email: balraj.13075@lpu.co.in Satoshi Takahashi Tsukuba University, Japan Yunqiong Wang Yunnan Normal University, China
Yosuke Motoki Yamagata University, Japan
Qinling Xin Central China University of Technology, China
Amandeep Nagpal Lovely Professional University, India Email: amandeep.nagpal@lpu.co.in
Shaochun Xu Algoma University, Canada Email: simon.xu@algomau.ca
List of Contributors Tianwei Xu Yunnan Normal University, China Email: xutianwei@ynnu.edu.cn Rongfang Yang Yunnan Normal University, China Xie Yongqiu Capital University of Economics and Business, China Hao Yu Capital University of Economics and Business, China Jinhui Yuan Renmin University of China Email: jcyjh@126.com Li Yunpeng Beijing University, China Kem Z.K. Zhang City University of Hong Kong, China Email: zikzhang@cityu.edu.hk
XV Jin Zhang Hainan University, China Email: zj001_cn@163.com Sesia J. Zhao USTC-CityU Joint Advanced Research Center, China Email: sesiazj@mail.ustc.edu.cn Hui Zhou Hainan University, China Hongwei Zhou Renmin University of China, China Email: hong wei zhou@hotmail.com Juxiang Zhou Yunnan Normal University, China Email: zjuxiang@126.com Xiaolin Zhou National University of Defense Technology, China Zongtan Zhou Peking University, Beijing Ge Zhu Beijing Yong You Software Company, China
Applying Bounded Model Checking to Verifying Web Navigation Model Honghao Gao, Huaikou Miao, Shengbo Chen, and Jia Mei*
Abstract. With the development of Web applications, formal verification of Web navigational behaviors has been a significant issue in Web engineering. Due to the features of Web technologies, such as caching, session and cookies, Web users can press the Back or Forward buttons to revisit Web pages. But these complex interactions between users and Web browsers may negatively influence the overall functionalities and navigations of Web applications. There are two challenges: One is that it is hard to model all possible navigation paths because the number of dynamic interactions and the personalized pages generated by different Web users may be huge or even infinite. Another is that how to improve the efficiency of verification because counterexamples usually manifest in a small number of navigation model. In this paper, to overcome these two challenges, we primarily present an On-the-fly approach to modeling the Web navigation behaviors, and apply bounded model checking (BMC) to verifying the On-the-fly navigation model. Finally, a prototype system is discussed.
1 Introduction Along with the appearance and rapid improvement of Internet, more and more companies are rushing to employ Web applications to support their business in order to accelerate the cooperation with their customers, i.e., B2B, C2C, G2C Web sites. Due to the distribution of Web environment, hyperlinks linked by page-topage paradigm have been used to binding the connection between Web pages. Moreover, the basic Web browser features provide an adequate set of navigational facilities for Web users to revisit Web pages, including the Back and Forward buttons, Refresh, Favorites, Link menu, URL-rewriting etc. Therefore, Web users can interact with not only the Web pages but also the Web browsers. However, Honghao Gao · Huaikou Miao · Shengbo Chen · Jia Mei School of Computer Engineering and Science, Shanghai University, 200072, Shanghai, P.R. China and Shanghai Key Laboratory of Computer Software Evaluating&Testing, 201112, Shanghai, P.R. China e-mail:{gaohonghao,hkmiao,schen,me269}@shu.edu.cn R. Lee (Ed.): Computer and Information Science 2011, SCI 364, pp. 1–15. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
2
H. Gao et al.
negatively pressing the Back or Forward buttons may influence the overall functionalities and navigations of Web applications, especially in Safety Critical Region (SCR)[3]. At present, how to modeling and verifying Web applications is still a complex task. To the best of our knowledge, the navigation model is one of the important research areas, also named navigation graph, which can help for clarifying requirements and specifying implementation behaviors. e.g., it is valuable for checking the conformance between a designed navigation behavior and an implemented navigation behavior. Actually, the navigation of a Web application is the possible sequence of Web pages a user has visited, where nodes represent Web pages and edges represent direct transitions between pages. The next page is determined by the current page and the action, i.e, back, forward, reload and hyperlink. But it also brings up two challenges: (1) the number of dynamic interactions and the personalized pages generated by different Web users may be huge or even infinite. It is nearly always impractical to model all possible navigation paths; (2) counterexamples only manifest in a small number of navigation model. It is inefficient to check all states of each path, many of which do not contribute to counterexample detection. In our paper, there are two novelties: (1) capturing dynamic navigation model on the fly and considering not only Web pages’ hyperlink but also Web browser buttons’ state; (2) using BMC to checking the properties in the small scope of navigation model for improving the efficiency of verification. Concretely, our approach involves three major steps: First, a Web Browser Loading Model (BLM) is used to construct the integrated model of the Web application incorporating the abstract behavior of the internal session control, caching mechanism and enabled buttons of web browser. Second, based on BLM, Kripke structure is employed to describe the On-the-fly navigation models by gradually combining each page’s navigation associations. Third, BMC is used to verify the safety and liveness property against the On-the-fly navigation model for enhancing the capacity of detecting faults. The remainder of this paper is organized as follows: Section 2 analyzes the Web applications behaviors, and introduces Kripke structure to construct the Web navigation model on the fly based on our previous study [3]. Section 3 gives a brief introduction to BMC, and then applies BMC to verifying the liveness and safety properties of Web applications. Section 4 reviews the related works. Section 5 draws a conclusion and future works.
2 Modeling Web Application Behaviors The purpose of this section is to discuss the features of Web browsers and Web pages for formally modeling the Web Application behaviors and the definitions of On-the-fly Web navigation model. The pages in SCR require higher security requirements. They prohibit the user from entering to SCR through pressing the Web button from no-SCR, even though the page has been visited in the past. As Fig.1 shown, we first introduce an
Applying Bounded Model Checking to Verifying Web Navigation Model
3
example of Audit System, which will be used throughout the paper. After receiving a request of login from a user, Audit System will authenticate his identity. Once he login system successfully, two things may happen: user can check out a bank receipt and audit the authenticity and reliability. Otherwise, user can declare this receipt invalid. During this process, the login page P2 can be linked to the set of SCR {P3, P4, P5, P7}. At page P4 or P7, the user can press the Back and Forward button to its predecessor and successor page, respectively. But at the boundary {P2, P3} or {P5, P6}, user can press the Web buttons to revisit pages{P3, P5} form pages{P2,P6}.Otherwise, the user will be confused to be redirected to an unreachable or unsecure Web page. To model these navigational behaviors of Web Applications, we should take the Web browser button click and the Web page hyperlink action into consideration.
Fig. 1 A motivation example
Actually, many Web browsers cache page contents. When a user triggered a series of operations, the Web browser will maintain a history stack for the revisited Web pages. The history stack consists of stack pointer, top pointer and bottom pointer specifying the button states. We mainly investigate the Replacement hyperlinks paradigm where the destination page replaces the source one[1]. Some of sophisticated situations [1], such as OpenInNewWindow and ShowInTooltipWindow, are not considered because our work only requires to obtain the available pages states. Definition 1 (Button enabled). If the history stack has more than one item and the stack pointer does not point to the bottom item stored, the Back button is enabled. Similarly, the Forward button is enabled when the history stack has more than one item and the stack pointer does not point to the top item [3]. Let n be the length of history stack. pno indicates the page position in the stack. For Back and Forward buttons, we consider four states [2]: (1) Back Disabled and Forward Disabled(BDFD), pno=1, n=1; (2) Back Disabled and Forward Disabled(BDFD), pno=1, n>1; (3)Back Enabled and Forward Disabled (BEFD), pno = n, n>1; (4)Back Enabled and Forward Enabled (BEFE), pno>1, pno m). By checking the original IP headers encapsulated in the ICMP replies, we are able to know an ICMP packet is returned by router at which hop. If rj responds but ri doesn’t (j > i ≥ 1), a router at hop i is loss. In this way, we find TC loses 2.3% routers and 3.3% links. Note that if rj isn’t the destination and all routers (or hosts) behind rj do not generate ICMP responses, we won’t be aware of the path behind rj. Sets S2 and S3 observe no more IP addresses (or routers) than S1, but 22 more links. By checking the geography location of small ISPs inside the ISP cloud and their IP address space, we find that all these 22 links are located at the border of the ISP cloud. Since most of the vantage points in S1 are located inside the ISP cloud, it is easy to assume that S1 may lose some edge links (and it does). But TC can not lose a considerable number of edge links because all 18 vantage points in S2 and S3 can only observe about 0.05% additional links. Nevertheless, these 22 links are added to TC. The latest datasets released by Skitter do not introduce new routers or links. Skitter detects only 9,093 IP addresses and 15,022 links of the ISP cloud. About 97.9% IP addresses are included by TC, but the left 2.1% IP addresses are unreachable (even during the validation). We suspect that since Skitter and we start probing at different time (12 days apart), Skitter happens to observe some routers and links that do not exist during our experiment. This indicates that TC is not an instantaneous topology of the underlying network; instead, TC is a snapshot of the ISP cloud over a time interval τ. If τ is too long, parts of the snapshot tend to be out-of-date. We argue that since the total time our vantage points take to probe and resolve IP alias is 70 hours, and since end-to-end route won’t change over time scales of days to weeks [15], TC is an accurate snapshot of the ISP cloud.
An Empirical Study of Network Topology Inference
221
In addition, TC includes six routers and 14 links that are outside the valid IP address space of the ISP cloud. Further investigation reveals that the USTC campus network has temporarily routed a small portion of its traffic through a local commercial network when one of its gateways breaks down. But these temporary routes are no longer available after the gateway is repaired. Since these six routers and 14 links are not part of the ISP cloud, they are cut out of TC. Furthermore, we ask the network operators of seven other campuses if they encounter similar situations during our experiment, and they all report no. We consult ten ISPs whose networks cover almost half of the ISP cloud, and they confirm that TC misses very few routers or links. The ISPs do not report specific ratios because their networks are so large that they do not have a complete map covering every corner except the backbones. But they claim that they are not aware of about 2% links, which connect to their backbone routers in TC. These links indeed exist since we can still detect them after the experiment. We suppose this situation is caused by local network operators who arbitrarily deploy fibers between backbone routers without reporting to their administrators immediately. We ask a question about how many routers in the ISP cloud are configured so as not to generate any ICMP packet. All ISPs answer that most of their routers can generate ICMP packets, and they also use ping or traceroute like toolkits for troubleshooting. So we are more confident with the result of the above selfverification on TC. Finally, we believe that TC captures most of the routers and their links of the ISP cloud.
4 Data Analysis After collecting and validating the topology information, we now evaluate the sampling bias by comparing TC with the aggregate topology built on a random set of vantage points in S1. The comparison was focused on topology coverage, metrics, and node degree distribution. All possible topologies that are built on the information of the 49 vantage points in S1 are arranged in 49 groups, G1, G2, ..., G49. The first group G1 consists of 49 topologies observed independently by the 49 vantage points. The second group G2 includes totally =1,176 topologies that are constructed by merging the topology information of every two different vantage points. Similarly, group Gx consists of topologies built on the information of every x different vantage points. Finally, G49 has only one topology which combines the views of all the 49 vantage points. Note that the topology in G49 is slightly different from TC, which has 22 additional links observed by S2 and S3, and discards six routers and 14 links that are outside the ISP cloud (see topology validation, Section 3). Fig. 4 shows the maximum and the minimum router coverage of the topologies in select groups G1, G5, ..., G49. For example, a topology in G1 can include at most 69% or at least 52% routers of TC. We find that even though vantage points in S1 are assigned the whole list of potential IP addresses of the ISP cloud, many
222
H. Zhou et al.
vantage points still fail to detect a large portion of routers. We check the topologies in G1 and find many routers in several ISPs are unreachable to the vantage points in some other ISPs due to AS-level policies, as also found in [23]. In addition, a router with the target IP address may respond to our probing packets through NICs that are assigned other IP addresses [12]. Naturally, as the number of vantage points increases, the number of routers they observe increases quickly as well. The “max” column is much higher than the “min” one when the index of group is less than 21. As the index continues to grow, the “min” column catches up gradually.
Fig. 4 The maximum and the minimum router coverage of topologies in Gx (x = 1, 5, ..., 49).
Fig. 5 The maximum and the minimum link coverage of topologies in Gx (x = 1, 5, ..., 49).
Fig. 5 plots the maximum and the minimum link coverage of topologies in G1, G5, ..., G49. In contrast to router coverage, all topologies in G1, G5, and G9 observe a very small portion of links. Moreover, when the index of group x is less than 21, different combinations of x vantage points obtain diverse link coverage. In fact, what a vantage point observes is a tree-like graph (not necessarily a tree). Particularly, the topologies in G1, G5, and G9 seem like a bundle of trees spread over TC. Many links are still undetected though the composite of a few trees indeed covers a large portion of TC. Furthermore, if a set of vantage points
An Empirical Study of Network Topology Inference
223
are logically close nodes in TC, it observes many common links and its link coverage would be comparatively low. In contrast, loosely connected vantage points often reach high link coverage because their views do not share many links. In the rest of this paper, the topology that achieves the highest link coverage in group Gx is termed Gx–Max, while the topology with the lowest link coverage in Gx is named Gx–Min.
5 Discussion What is the cause of sampling bias? The sampling bias is mainly determined by topology coverage. The percentage of target network that the inferred topology covers can strongly bias the observations of the target network. Furthermore, the number of vantage points and their locations will significantly affect the topology coverage. Finally, an inferred topology is just a snapshot of real network. So the time period during which a network is measured should be as short as possible, otherwise the network would undergo considerable changes and the inferred topology is prone to inaccuracy or incompleteness. Note that this paper focuses largely on link coverage, but the link coverage is enough to tell the importance of topology coverage in sampling bias. What does the sampled information tell us about the real network? Here, the sampled information refers to the inferred topology, which is sampled because it is usually impossible to obtain a complete and instantaneous picture of target network. Therefore, we have to characterize the network using sampled information, and we suspect that the sampled information can tell us any possible information about the target network. Though the study of an inferred topology would also lead to the same conclusions on a few properties as one does with a real topology, it may not be very safe to assume other properties of the inferred topology match those of the real one. How to capture an accurate topology with as few measurements as possible? Despite the challenges of mapping networks, it is possible to capture the accurate topology of a target network with a small number of vantage points. The prerequisite is that these vantage points should be placed in suitable locations of target network in order to achieve high all-point-distance. To do so, we need to make careful trade-offs between topology coverage and measurement time. First, to maximize the topology coverage, the all-point distance of a fixed number of vantage points in all available positions must be computed so as to find suitable locations. In addition, an appropriate probing strategy is necessary. What is an accurate topology, anyway? The answer to this question is metricspecific, meaning that it depends on which metric is under estimation. For example, in our analysis, G13–Max is accurate if only distortion is taken into account, but it is not accurate as resilience is involved. In addition, the answer also varies with the required exact level of metrics. For example, if our purpose is to check whether or not the node degree distribution is a power law rather than calculate the parameters of distribution precisely, most of the inferred topologies seem accurate.
224
H. Zhou et al.
Therefore, to obtain accurate estimation of metrics from a comprehensive perspective, an accurate topology should be the topology that achieves high topology coverage. But how much coverage can be regarded as “high” coverage depends on the required exact level of the metrics that we are interested in.
6 Conclusions Understanding the sampling bias is very important because it enables us to link an inferred topology to the real network reasonably. This paper systematically evaluates the sampling bias of network topology inference. Our basic idea is to compare inferred topologies with an almost complete topology of a specific and large-scale network from various perspectives. To do so, we identify an ISP cloud, spread vantage points over the ISP cloud and the world, collect topology information by probing a fixed list of IP addresses, merge the views of all vantage points to produce the almost complete topology, which consists of 25,733 routers and 36,029 links, and validate this topology. We find that sampling bias, if undetected, could significantly undermine the conclusions draw on the inferred topologies. Moreover, an inferred topology that shares the same properties of target network may still be thought inaccurate if other properties are involved. Finally, sampling bias is associated with topology coverage (especially link coverage) that the inferred topology can achieve. To weaken the effect of sampling bias, researchers should carefully select the geography location of vantage points so as to achieve high all-point-distance, focusing on specific metrics, and predict the scope of target network before measurement starts.
Acknowledgment We gratefully acknowledge the financial support of the Project 211 supported coordinately by the State Planning Commission, Ministry of Education and Ministry of Finance, China.
References [1] Meyer, D.: Routeviews, http://www.routeviews.org/ [2] Spring, N., Mahajan, R., Wetherall, D., Anderson, T.: Measuring ISP topologies with Rocketfuel. IEEE/ACM Trans. Networking 12(1), 2–16 (2004) [3] Broido, A., Claffy, K.: Internet topology: connectivity of IP graphs. In: Proc. SPIE ITCom WWW conference, August 2001, pp. 172–187 (2001) [4] Burch, H., Cheswick, B.: Mapping the Internet. IEEE Computer 32(4), 97–98 (1999) [5] Govindan, R., Tangmunarunkit, H.: Heuristics for Internet map discovery. In: Proc. IEEE INFOCOM, pp. 1371–1480 (2000)
An Empirical Study of Network Topology Inference
225
[6] Breitbart, Y., Garofalakis, M., Jai, B., Martin, C., Rastogi, R., Silberschatz, A.: Topology discovery in heterogeneous IP networks: the NetInventory system. IEEE/ACM Trans. Networking 12(3), 401–414 (2004) [7] Kernen, T.: traceroute organization, http://www.traceroute.org/ [8] Cooperative Association for Internet Data Analysis (CAIDA): http://www.caida.org/ [9] Faloutsos, C., Faloutsos, P., Faloutsos, M.: On power-law relationships of the Internet topology. In: Proc. ACM SIGCOMM, September 1999, pp. 251–262 (1999) [10] Chen, Q., Chang, H., Govindan, R., Jamin, S., Shenker, S., Willinger, W.: The origin of power laws in Internet topologies revisited. In: Proc. IEEE INFOCOM (2002) [11] Willinger, W., Govindan, R., Jamin, S., Paxson, V., Shenker, S.: Scaling phenomena in the Internet: critically examining criticality. Proc. National Academy of Sciences 99(suppl.1), 2573–2580 (2002) [12] Paxson, V.: Measurements and analysis of end-to-end Internet dynamics. Ph.D. dissertation, Univ. California, Berkeley (1997) [13] Lakhina, A., Byers, J.W., Crovella, M., Xie, P.: Sampling biases in IP topology measurement. In: Proc. IEEE INFOCOM, pp. 332–341 (2003) [14] Abilene Network., http://www.internet2.edu/abilene/ [15] Zhang, Y., Duffield, N., Paxson, V., Shenker, S.: On the constancy of Internet path properties. In: Proc. ACM SIGCOMM conference on Internet measurement, pp. 197–211 (2001) [16] Floyd, S., Paxson, V.: Difficulties in simulating the Internet. IEEE/ACM Trans. Networking 9, 392–403 (2001) [17] China Education and Research Network (CERNet),, http://www.edu.cn/HomePage/english/cernet/index.shtml [18] Resilient Overlay Networks (RON), http://nms.lcs.mit.edu/ron/ [19] PlanetLab, http://www.planet-lab.org/ [20] Prtraceroute, http://www.isi.edu/ra/RAToolSet/prtraceroute.html [21] Postel, J.: Internet control message protocol. IETF, RFC 792 (1981) [22] Zhou, H., Wang, Y.: RichMap: combining the techniques of bandwidth estimation and topology discovery. Journal of Internet Engineering 1(2), 102–113 (2007) [23] Tangmunarunkit, H., Govindan, R., Shenker, S., Estin, D.: The impact of policy on Internet paths. In: Proc. IEEE INFOCOM (2001)
Computer Network Reverse Engineering Hui Zhou, Wencai Du, Shaochun Xu, and Qinling Xin
*
Abstract. Software reverse engineering has undergone many milestones and stepped from research to industry quickly in recent ten years. By analogy, we have found that it is also possible to apply reverse engineering to computer networks. The goal of network reverse engineering is to annotate a map of the networks with properties such as node distribution, connectivity, and bandwidth usage. It is necessary, but also challenging, to employ reverse engineering to computer networks. To do this, we first comparatively analyze the reverse engineering of both software and network from five basic perspectives: source, data analysis, presentation, validation, and prediction. And then, RichMap system has been developed to automatically infer the topology and link available bandwidth of a network. The experiment result indicates that, after applying the object snapshot concept of software, RichMap can smoothly capture and present complete router-level snapshots and significantly decrease the network load that it generates.
1 Introduction Software reverse engineering is, in practice, one of the most important endeavors in software engineering. This stems from the fact that software systems are complex and often poorly specified and documented. As a result, software practitioners need to spend a substantial amount of time understanding the source code from a structural and behavioral perspective, before carrying out any maintenance task. In this context, most reverse engineering processes follow the same pattern: a program is analyzed through static or dynamic analysis and the collected low-level program information is transformed into a higher level, more abstract presentation. The presentation helps engineers understand the rationale of the code and thus facilitate future refactoring. Given the dynamic nature of the Internet, keeping track of network information manually is a daunting (if not impossible) task. Network operators generally can’t Hui Zhou · Wencai Du Hainan University, Renmin Ave. No. 58, 570228, Haikou, China e-mail: wencai@hainu.edu.cn
*
Shaochun Xu Algoma University, Sault Ste, Marie, Ontario, P6A2G4, Canada Qinling Xin Central China University of Technology, Wuhan, China R. Lee (Ed.): Computer and Information Science 2011, SCI 364, pp. 227–239. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
228
H. Zhou et al.
draw a complete map of their networks since many internal parts can undergo different scales of changes but will not report these changes immediately. Therefore, they need reverse engineering systems to detect underutilized and congested links, plan network capacity upgrades, and deploy security infrastructure. In addition, many users also need to verify whether they get the network service stated in their service-level agreements with the Internet service providers (ISPs). It has become obviously necessary to employ reverse engineering to computer networks. However, network reverse engineering is a challenging task. The key reason is that the design of the Internet can’t provide explicit support for end nodes to obtain information about the network internals. A network typically consists of many small networks; such networks are under different administrative control, so there is no single place from which one can obtain a complete picture of the specified target network. Furthermore, the Internet is so heterogeneous that an approach found to be useful in a certain networks may not be effective elsewhere [1]. This paper makes two contributions. First, we analyze the differences between software reverse engineering and network reverse engineering from five basic perspectives: source, data analysis, presentation, validation, and prediction. In addition, we build RichMap system, which need to be installed on a single client host, to characterize and monitor its surrounding computer networks. The experiment result proves that, after adopting the snapshots concept from software domain, RichMap is able to present a series of router-level views of a large-scale network. And it can also effectively illustrate the changes of topology, congested links, and delay without injecting noticeable probing packets into target network. This paper is organized as follows. Section 2 summarizes the related work on network reverse engineering domain. Section 3 analyzes the reverse engineering techniques of both software and networking, and then Section 4 presents RichMap, which draws a series of streaming snapshots about designated networks. Section 5 discusses our findings, and finally Section 6 concludes the paper.
2 Related Works The field of software reverse engineering and its closely related fields, such as program comprehension or software analysis, have undergone many successes over the past 20 years. In addition, software reverse engineering environment has been equipped with various intelligent tools: extractors, analyzers, and repositories [2]. During the same time, along another thread, network community has introduced quite a few measurement systems to gathering and presenting the information of network properties [3]. The theories, protocols, techniques, tools, overlay framework, and the released data archives have initially make up the main body of network reverse engineering. Basically, the reverse engineering of network mainly starts from measurement. Specifically, a router can be configured to passively record the information about its own performance, e.g. the number of packets received/sent by each of its network interface cards (NICs). A typical example is network traffic monitoring. Fig. 1 illustrates the bytes sent through the USENET bulletin board system, averaged over two-week intervals.
Computer Network Reverse Engineering
229
Fig. 1 USENET traffic monitoring information [4].
Furthermore, the measurement literature can further be classified according to different measurement targets: node, link, topology. Learning the role that a node plays is the first step to understand the network. Basically, each node has one of the following roles: client host; access router that aggregates the traffic from clients; and backbone router that transmits a large volume of traffic. The role problem has been frequently addressed, e.g. Rocketfuel [5] uses IP prefixes, DNS information, and topological ordering to identify role. In addition, many tools search for the bottleneck node with diverse heuristics [6]. In addition to role, the behavior of node has been a key reverse engineering target. For example, TCP features and supporting network services both can affect the composition of traffic of a node [7]. Practically, a node may hold multiple NICs, each with a different IP address. To provide a reasonable node-level, instead of IP-level, network analysis, we must decide which interface belongs to the same node [5]. Besides node, link is another important component. Generally, a link is the IP connection between two nodes that are only one IP-hop away from each other. Much research has been done to capture the usability, delay, and bandwidth capacity of a single link. Recently, the research community extends the study of link to end-to-end path, which can be regarded as a line of connected links. Measuring the properties of a path is very meaningful since it enables us to better understand how packets flow between nodes. Typically, tools use Internet control message protocol (ICMP) [8] timestamps to estimate the delay variation. In addition to delay, the available bandwidth of path has attracted much attention since 1990s. Specifically, the available bandwidth is defined as the maximum rate that a path can provide to a packet flow, without reducing the rate of other flows in the path [9]. Measuring the instantaneous end-to-end available bandwidth is extremely difficult. We have examined 11 well-known available-bandwidth
230
H. Zhou et al.
measurement tools, and found that quite a few basic problems, e.g. system timing and end-host throughput, which can always lead to different scales of bias [1]. Finally, topology auto-discovery has strongly driven the study of active probing measurement. Network community has examined five categories of topologies: the graphs of connections between autonomous systems (ASs) [5], the point-ofpresence (POP) topologies that interpret the structure of backbone using geography information, the IP-level topologies whose nodes are IP addresses and whose links are connections between the IP addresses, the router-level topologies that resolve IP aliases and group the IP addresses in the unit of router, and the connectivity of physical components.
Fig. 2 The discovered topology of Abilene backbone [11].
For example, Breitbart et al. detected 3,888 nodes and 4,857 links in 2003 [3]. RocketFuel outputted a topology consisting of 228,263 nodes and 320,149 links in 2004 [5]. An ongoing project, Skitter, has been scanning the whole Internet for several years with tens of commercial network hosts, and it has released extensive graphs of Internet IP-level topologies [10]. As an example, Fig. 2 gives the result of a topology discovery work; the target network is Abilene backbone [11].
3 Comparative Analysis We comparatively analyzed the reverse engineering of software and network from five basic perspectives: source, analysis, presentation, validation, and prediction.
3.1 Source The source of software reverse engineering is code and code-related files such as log. Generally, software reverse engineering depends on performing some analysis of the source code in order to produce one or more models of the system under
Computer Network Reverse Engineering
231
analysis. Generally, source code is written by software engineers according to the well-designed specification of programming languages, e.g. ASM, Pascal, C/C++, and Java. A language often comes with a specification, to which compiler developer and software engineer must conform. Furthermore, the coding process is supported by various integrated development environments. As a result, no matter how well (or bad) the code is organized, software reverse engineering tools is built on a solid basis, i.e. the tools do understand the exact meaning of each line of code. Unlike the source of software reverse engineering, the one of network reverse engineering mainly comes from measurement, and it is highly volatile. The volatility can be perceived in almost every parameter that we attempt to measure. For example, the round-trip time (RTT) of a pair of nodes is an important metric of network performance. Generally, RTT can be used as an indicator of end-to-end transmission quality. Here we attempt to measure the RTT of a short path, i.e. two directly connected computers C1 and C2. First, C1 sends an ICMP echo-request packet to C2. When C2 receives the packet, it immediately sends an ICMP echo-reply packet back to C1. In each active probe, the time from sending out an ICMP echo-request to receiving the corresponding echo-reply is regarded as a candidate of RTT. As shown in Fig. 3, the RTT is ever-changing with network traffic and time.
Fig. 3 Round-trip time of two directly connected computers.
232
H. Zhou et al.
3.2 Analysis To analyze the source code, a software reverse engineering tool will first scan the source code. In most cases, reverse engineering tool assumes that the target source files won’t undergo any change during the scan, which is done once and for all. In a very limited time interval, the source of software is safe to be regarded as static, while network is always a moving target. As a result, network tools must continuously collect the information about the designated network, in a never-ending style. Moreover, as to network reverse engineering, analyzing the data source is challenging since it generally contains too much noises. But the analysis is valuable since it often provide insight into the network. For example, Faloutsos et al. discover some surprisingly simple power-laws of the network topologies [12]. These power-laws hold for three topologies between November 1997 and December 1998, despite a 45% growth of its size during that period. As shown in Fig. 4, loglog plot of the out-degree dv versus the rank rv in the sequence of decreasing out-degree.
Fig. 4 The rank plots on dataset Intel-98 [12].
3.3 Presentation Suppose that the presentation of software reverse engineering is a snapshot, the one of network reverse engineering can be regarded as a video. The parameters of target network can undergo changes as time passes, and thus lead to high dynamics. As shown in Fig. 5, the IP conversations of LAN captured by Sniffer Pro, which is a network packet sniffing tool installed in one node [13]. Since the target network is ever-changing, the presentation must trace the changes and output pictures that match.
Computer Network Reverse Engineering
233
Compared with software reverse engineering, the network reverse engineering tools can’t support large-scale reuse since there isn’t a universal accepted presentation standard. It is also hard to establish such a standard because each reverse engineering tool is built to study a specific question and work in a specific network environment.
Fig. 5 IP conversations captured by Sniffer Pro in 9:00 – 9:06 PM.
3.4 Validation To validate the available bandwidth of a path, researchers have introduced many inspiring techniques. It seems that comparing the estimation result with closely estimated bulk TCP throughput over the same path is a good idea [14]. However, available-bandwidth and bulk TCP throughput are indeed different. The former gives the total spare capacity in the path, independent of which transport protocol attempts to capture it. While the latter depends on TCP’s congestion control. Fig. 6 typically shows the measurement result of the available bandwidth of an end-toend path, which starts from Hainan University and ends at Chinese Academy of Sciences. In particular, Cprobe [15] and BNeck [6] are installed on hosts inside Hainan, Pathload [16] is installed in both end points, while TCP throughput is tested by maximized the parallel TCP connections of Iperf [17]. It is apparent that there isn’t a curve that can exactly match the other. As a result, we are not able to completely validate end-to-end available bandwidth. Furthermore, it is very hard to make sure the data we collect reflects the exact network status, even if we have success experience on a limited number of networks. The same problem is faced by almost all measurement techniques that rely on active probing. And this thus makes the network reverse engineering more challenging than its software counterpart.
234
H. Zhou et al.
Fig. 6 Available bandwidth measured by different tools.
3.5 Prediction Recently, there is a growing need of reverse engineering tools to support the prediction of changes in source. For example, through analyzing the history of the lines of code, managers can predict the code scale of a Java program in the next development iteration [2]. Surprisingly, though network contains much more noise than stationary software source code, many useful rules have been extracted, and used to predict the macro-behavior of networks. Diurnal patterns of activity: It has been recognized for more than thirty years that network activity patterns follow daily patterns, with human-related activity beginning to rise around 8-9AM local time, peaking around 11AM, showing a lunch-related noontime dip, picking back up again around 1PM, peaking around 34PM, and then declining as the business day ends around 5PM. The pattern often shows renewed activity in the early evening hours, rising around say 8PM and peaking at 10-11PM, diminishing sharply after midnight. Originally, this second rise in activity was presumably due to the “late night hacker” effect, in which users took advantage of better response times during periods of otherwise light traffic load. Self-Similarity: Longer-term correlations in the packet arrivals seen in aggregated Internet traffic are well described in terms of self-similar processes [18]. “Longer-term” here means, roughly, time scales from hundreds of milliseconds to tens of minutes. The traditional Poisson or Markovian modeling predicts that longer-term correlations should rapidly die out, and consequently that traffic observed on large time scales should appear quite smooth. Nevertheless, a wide body
Computer Network Reverse Engineering
235
of empirical data argues strongly that these correlations remain non-negligible over a large range of time scales. While on longer time scales, non-stationary effects such as diurnal traffic load patterns (see previous item) become significant. On shorter time scales, effects due to the network transport protocols, which impart a great deal of structure on the timing of consecutive packets, appear to dominate traffic correlations [19].
4 RichMap To start network reverse engineering, and to accurately capture the running status of a network, we developed an experimental system: RichMap [21]. RichMap has three basic features. First, it is built on active probing technique, and is a singlenode system instead of an overlay network system like Planet-Lab [22] that requires software to be installed on many nodes. Second, it automatically discovers the node-level topology of surrounding network, as well as link available bandwidth and delay variation. Finally, it utilizes the snapshot concept from software domain, and builds series of easy-to-understand network maps smoothly. From boot time, RichMap starts a process to continuously measure the target network. When the RichMap is requested, it presents a map. If the request happens after the end of a measurement cycle and before the start of a new cycle, RichMap updates the repository with the information collected in the latest cycle. But, most of the time, the request occurs during the course of current cycle. At this time, RichMap displays the reverse engineering result of current cycle over the map of the last cycle, while the nodes and links of old map (judged by timestamp) are shadowed. To evaluate RichMap, we installed it on a node that was in the same LAN of a backbone router in Tsinghua University, and configured RichMap to reverse engineer the network of teaching building No. 3. Fig. 7 gives the 54th and 60th hour snapshots of the outputted map. We observed that there were about ten high-speed links, connecting many local networks. About eight networks were built with high-performance equipments, while many others were not. It was also valuable to note that only the nodes with public IP addresses were drawn, a large number of nodes owned by individual department and accessed the Internet through network address translation technology were not included. We also found that the available bandwidth of backbone links was steady, while the available bandwidth of non-backbone links tended to fluctuate. Link available bandwidth of the 54th-hour snapshot was generally higher than that of the 60th-hour one. The reason was that the 54th-hour snapshot was collected at night, while the 60th-hour one was in the morning.
236
H. Zhou et al.
Fig. 7 Snapshots outputted by RichMap at 54th and 60th hour.
Besides the smooth presentation effect, adopting the snapshot idea could significantly decrease the network load. As shown in Fig. 8, when RichMap closed the snapshot option, it needed to actively probe the network one cycle by another. When the option was open, RichMap could pause a while in between two adjacent cycles. This was very useful especially when we choose to reverse engineering the
Computer Network Reverse Engineering
237
network at a specific time, and we found the number of nodes discovered by RichMap, no matter it turned on the option or not, were almost the same (Fig. 9).
Fig. 8 End-host throughput.
Fig. 9 The number of nodes detected by RichMap.
5 Conclusions Reverse engineering is the process of studying the design of an object from its implementation. Reverse engineering has long rooted in software field, and now we found it useful to promote creative applications for the computer networks. A typical sample is the RichMap system; the snapshot concept enables it to present a series of steady maps of target network. With RichMap, we discuss the possibility and benefit of network reverse engineering, and argue that the reverse engineering is within the reach of both software and network communities.
238
H. Zhou et al.
Acknowledgment We gratefully acknowledge the financial support of the Project 211 supported coordinately by the State Planning Commission, Ministry of Education and Ministry of Finance, China.
References [1] Zhou, H., Wang, Y., Wang, X., Huai, X.: Difficulties in Estimating Availablebandwidth. In: Proceedings of IEEE International Conference on Communications, pp. 704–709 (2006) [2] Kienle, H.: Building Reverse Engineering Tools with Components. Ph.D. Thesis, Department of Computer Science, University of Victoria, Canada; 325 p (2006) [3] Breitbart, Y., Garofalakis, M., Jai, B., Martin, C., Rastogi, R., Silberschatz, A.: Topology Discovery in Heterogeneous IP Networks: the NetInventory System. IEEE/ACM Trans. Networking 12(3), 401–414 (2004) [4] Thompson, K., Miller, G., Wilder, R.: Wide-area Internet Traffic Patterns and Characteristics. IEEE Network, 10–23 (1997) [5] Spring, N., Mahajan, R., Wetherall, D., Anderson, T.: Measuring ISP Topologies with Rocketfuel. IEEE/ACM Trans. Networking 12(1), 2–16 (2004) [6] Zhou, H., Wang, Q., Wang, Y.: Measuring Internet Bottlenecks: Location, Capacity, and Available Bandwidth. In: Proceedings of International Conference on Computer Network and Mobile Computing, pp. 1052–1062 (2005) [7] Padhye, J., Floyd, S.: Identifying the TCP Behavior of Web Servers. In: Proceedings of. ACM SIGCOMM (2001) [8] Postel, J.: Internet Control Message Protocol. IETF RFC 792 (September 1981) [9] Dovrolis, C., Ramanathan, P., Moore, D.: Packet Dispersion Techniques and a Capacity Estimation Methodology. IEEE/ACM Trans. Networking 12, 963–977 (2004) [10] Cooperative Association for Internet Data Analysis (CAIDA), http://www.caida.org/ [11] Abilene Network, http://www.internet2.edu/abilene [12] Faloutsos, M., Faloutsos, P., Faloutsos, C.: On Power-law Relationships of the Internet Topology. In: Proceedings of ACM SIGCOMM, Cambridge, USA (1999) [13] Sniffer Pro, http://www.netscout.com/ [14] He, Q., Dovrolis, C., Ammar, M.: On the Predictability of Large Transfer TCP Throughput. Computer Networks 51(14), 3959–3977 (2007) [15] Carter, R., Crovella, M.: Measuring Bottleneck Link Speed in Packet-switched Networks. Performance Evaluation 27(28), 297–318 (1996) [16] Jain, M., Dovrolis, C.: End-to-end Available Bandwidth: Measurement Methodology, Dynamics, and Relation with TCP Throughput. IEEE/ACM Trans. Networking 11(4), 537–549 (2003) [17] Tirumala, A., Qin, F., Dugan, J., Ferguson, J., Gibbs, K.: Iperf - The TCP/UDP Bandwidth Measurement Tool, http://dast.nlanr.net/Projects/Iperf/
Computer Network Reverse Engineering
239
[18] Zhang, Y., Duffield, N., Paxson, V., Shenker, S.: On the Constancy of Internet Path Properties. In: Proceedings of ACM SIGCOMM conference on Internet measurement, pp. 197–211 (2001) [19] Paxson, V.: End-to-end Internet Packet Dynamics. In: Proceedings of ACM SIGCOMM (1997) [20] Yuvrai, A., et al.: Somniloquy: Augmenting Network Interfaces to Reduce PC Energy Usage. In: Proceedings of the 6th USENIX Symposium on Networked Systems Design and Implementation (2009) [21] Zhou, H., Wang, Y.: RichMap: Combining the Techniques of Bandwidth Estimation and Topology Discovery. Journal of Internet Engineering 1(2), 102–113 (2008) [22] Turner, J., et al.: Supercharging Planetlab: a High Performance, Multi-application, Overlay Network Platform. ACM SIGCOMM Computer Communication Review 37(4), 85–96 (2007) [23] Gkantsidis, C., Karagiannis, T., Vojnovi, M.: Planet Scale Software Updates. ACM SIGCOMM Computer Communication Review 36(4), 423–434 (2006)
CUDA-Based Genetic Algorithm on Traveling Salesman Problem Su Chen, Spencer Davis, Hai Jiang, and Andy Novobilski
Abstract. Genetic algorithm is a widely used tool for generating searching solutions in NP-hard problems. The genetic algorithm on a particular problem should be specifically designed for parallelization and its performance gain might vary according to the parallelism hidden within the algorithm. NVIDIA GPUs that support the CUDA programming paradigm provide many processing units and a shared address space to ease the parallelization process. A heuristic genetic algorithm on the traveling salesman problem is specially designed to run on CPU. Then a corresponding CUDA program is developed for performance comparison. The experimental results indicate that a sequential genetic algorithm with intensive interactions can be accelerated by being translated into CUDA code for GPU execution.
1 Introduction Genetic algorithm (GA) and other stochastic searching algorithms are usually designed to solve NP-hard problems [3]. The traveling salesman problem (TSP) is a famous NP-hard problem [5][10]. It aims to get the shortest wraparound tour path for a group of cities. Since NP-hard problems cannot be solved in acceptable time, people aim to find acceptable solutions in acceptable time instead. To achieve this, various heuristic algorithms, such as the genetic algorithm, ant algorithm, tabu search, neural network, etc., are designed. Genetic algorithm was inspired by the evolvement of chromosomes in the real world, which includes crossover, mutation, and natural selection. Viewing chromosomes as solutions to a TSP problem, crossover and mutation are Su Chen · Spencer Davis · Hai Jiang · Andy Novobilski Department of Computer Science Arkansas State University, Jonesboro, AR, 72467, USA e-mail:{su.chen,spencer.davis}@smail.astate.edu, {hjiang,anovobilski}@astate.edu R. Lee (Ed.): Computer and Information Science 2011, SCI 364, pp. 241–252. c Springer-Verlag Berlin Heidelberg 2011 springerlink.com
242
S. Chen et al.
changing phases for chromosomes, while natural selection is a sifting phase that will wash out the worst solutions so that better ones will stay. To simulate this process in a computer program, programmers have to design sequences of numbers to represent chromosomes and perform certain operation on them. Different Problem will have different types of chromosome designs. For example, select participants from a group will have a design of a 1a 2a 3 ...an as its chromosome, where ai are either 0 or 1 , where 0 means unselected and 1 means selected. As the problem size increases, it takes a very long time to reach an optimum solution, or even a less-optimum but satisfying solution. In order to shorten the convergence time, artificial intelligence is usually introduced to make algorithms efficient. For the traveling salesman problem, 2-opt is a specifically designed mutation operator which takes longer time than ordinary operators, but guarantees fast and steady convergence. However, even with this efficient operator, computing time is still quite long when problem size is large. Recently, NVIDIA’s CUDA programming paradigm enables GPU as a new computing platform [1][2]. Many-core GPUs can explore parallelism inside Genetic Algorithms for execution speedup and provide a cost effective method of implementing SIMD type solutions. This paper intends to develop a heuristic genetic algorithm on TSP and then parallelize it with CUDA on GPUs for performance gains. The rest of the paper is organized as follows: Section 2 discuss the deployment of genetic algorithm on TSP problem. Section 3 addresses the issues of genetic algorithm implementation on GPUs with CUDA. Section 4 provides performance analyses on both CPU and GPU. Section 5 gives the related work. Finally, our conclusions and future work are described.
2 Genetic Algorithm on TSP GA’s input usually includes a waypoint number and a distance table. The Output of GA should be an optimized chromosome chain that represents the order of cities that the traveling salesman should follow. The general process of GA is given in Fig. 1. The Initialization phase generates a group of chromosomes as shown in Fig. 2. The group size can influence quality of the final result and running time. Therefore, it needs to be properly chosen. Generally, when the group size increases, results are potentially better whereas the running time increases. Factors for both good results and a reasonable running time should be considered. Crossover phase is an important part in GA to simulate the action where two chromosome individuals exchange partial sections of their bodies. This process helps increase diversity as well as exchange better genes within population. Crossover on real chromosomes is illustrated in Fig. 3. Unfortunately, in TSP, there are no two same numbers in one chain. Therefore, it is
CUDA-Based Genetic Algorithm on Traveling Salesman Problem
243
Initialization
Crossover
Mutation
Selection
NO Reach condition?
YES Output results
Fig. 1 Flow chart of the general process in Genetic Algorithm (GA)
0
1
5
2
4
3
0-2-4-3-5-1-0
Fig. 2 Initial chromosome sequence generated from Genetic Algorithm (GA)
A
G
T
G
C
A
G
A
A
T
C
C
A
G
C
G
T
G
A
T
C
Exchange real sections C
A
T
C
A
T
C
Fig. 3 A crossover example with actual exchange in the real world
impossible to do crossover directly, as chromosomes do in the real world. However, there are alternative ways to simulate this process. The strategy used by this paper is based on sequence orders not values, as shown in Fig. 4 where the crossover of the selected portion of chromosomes is reasonably done.
244
S. Chen et al.
0
2
1
3
4
{2nd, 1st, 3rd}
0
2
1
3
1
2
3
1
2
3
5
0
0
1
2
4
3
5
0
{1st, 2nd, 3rd} Exchange order information {2nd, 1st, 3rd}
{1nd, 2nd, 3rd}
4
5
0
0
{1st, 2nd, 3rd}
1
2
4
2
1
4
2
1
4
{1nd, 2nd, 3rd}
{2nd, 1st, 3rd}
3
5
0
{2nd, 1st, 3rd}
Fig. 4 The crossover in the GA on Traveling Salesman Problem (TSP)
Since good chromosomes are forced to stay in population and pass down their heritage information by crossover, after generations, chromosomes will assimilate each other. The mutation phase is designed to make unpredictable changes on chromosomes in order to maintain the variety of the population. Mutation operators can be arbitrarily designed but the effects taken by them will be hard to tell. Some mutation operators will slow down the convergence process, while others will accelerate it. In this paper, we select 2-opt as the mutation operator, which can make the algorithm converge much faster than ordinary GA. The 2-opt mutation operator is specifically designed to solve TSP and guarantees both diversity and steady evolvement [5][10]. However, this operator takes O(n) time, and has larger time cost than that of simple operators. Details in 2-opt is given in Fig. 5.
0
0
1
2
5
1
4
2
5
4
3
3
0-1-2-4-3-5-0
0-2-4-3-5-1-0 0 5
1
2
4
3
0-1-2-3-4-5-0
Fig. 5 One possible mutation example in Genetic Algorithm
CUDA-Based Genetic Algorithm on Traveling Salesman Problem
245
Selection phase is usually placed after crossover and mutation. In this paper, simplest selection method is adopted and only better solutions are accepted. Each new chromosome will be compared with the older one and the better of the chromosomes stay in the population. A termination condition should be set to stop the evolution process. In this paper, when the best result has stopped evolving for some generations, algorithm will stop and output the result. Though better solutions are expected when execution time becomes longer, after solutions are convergent, the probability for GA to update the best result becomes extremely small.
3 Genetic Algorithm Implementation with CUDA 3.1 CUDA Platform and GPU Architecture CUDA (Compute Unified Device Architecture), developed by NVIDIA, is a parallel programming paradigm [1][2]. While graphics cards were originally designed only to process image and video flows, CUDA provides a platform to solve any general purposed problem on GPU. Rather than fetching image pixels concurrently, now threads in the GPU can run common tasks in parallel; however, as in other parallel programming platforms, task dependency problems should be considered by programmers themselves. Besides hundreds of threads, Fermi (The latest GPU Architecture in 2010) provides shared memory that can be accessed by threads within the same block extremely fast. Shared memory can be thought of as a cache that can be directly manipulated by users. When the input of the problem is small and all its intermediate results can be loaded into shared memory, Fermi will do excellent job. On the other hand, if the input size is relatively large, the utilization of shared memory should be carefully considered. Limitations of CUDA cannot be ignored. Recursion and pointers for functions are still not supported and debugging is a tedious job. The bus latency between the CPU and GPU exhibits as a bottleneck. All of these limitations should be avoided or considered during programming, and CUDA’s architecture should be taken advantage of in their code.
3.2 Opportunities for GA with CUDA Usually, in order to guarantee the diversity of species, GA maintains a group, or population, consisting of a good number of chromosomes. This can be thought of as the desire of the problem solver to create more directions in order to search a bigger area. It can be understood as that if more ants are dispatched to different directions, chance to find food becomes greater. In GA, these ants communicate with each other frequently and change their
246
S. Chen et al.
searching directions based on the information they get. Though it contains many interactions and dependencies, it is possible to be parallelized for performance gains. The most reasonable way to parallel this process is to map activities of chromosome individuals to separated threads. Since all chromosomes will do the same job, they roughly finish at the same time. This property prevents cores from being idle. Otherwise, synchronization will drag down performance severely. CUDA platform and Fermi architecture provide good tools to parallelize the algorithm. First, GPU supplies hundreds of cores for executing threads in parallel. Second, threads can talk to each other easily and fast because they share address space in several levels such as shared memory and global memory levels. Third, as an extension of C, CUDA eases the programming task.
3.3 Random Number Generation in CUDA Since GA is a stochastic searching algorithm, a random numbers generation strategy is required. Unfortunately, CUDA does not provide one yet. However, a pseudo random number generator can be easily simulated in different ways. In this paper, bit shifting, multiplication and module operators are used to generate random numbers. CUDA programs on GPUs may slow down when threads compete for random seeds. To solve this problem, random seeds are generated in the CPU and assigned to each GPU thread. Equipped with a simple random number generator function, threads in the GPU can generate random numbers simultaneously without blocking or false sharing.
3.4 Data Management for GA In CUDA architecture, threads can be arranged in blocks and grids to fit applications. In the latest Fermi architecture, cache and shared memory coexist to enable GPU cores to behave as CPU. This provides a greater chance to get better performance. Shared memory is as fast as cache and can be directly manipulated by users. However, shared memory space can only be accessed by threads within one block. If shared memory is big enough for everything, programmers do not have to spend too much time on data manipulation. However, with limited shared memory size, only few frequently used variables and arrays have priorities to reside in it. For the GA on TSP problem, the distance table and chromosome group occupy the majority of the space and both are too large for shared memory. Since distance between two cities is Euclidean distance in this paper, the distance table can be discarded and coordinate arrays are used instead. This change will definitely harm CPU’s performance because of duplicated
CUDA-Based Genetic Algorithm on Traveling Salesman Problem
247
calculations. However in GPU, such computation redundance is encouraged since there always are many idle cores due to memory access latency. This design is proved to be valid by experimental results.
3.5 Parallelization of GA Because each thread has an independent seed for random number generation, different threads can initialize chromosomes simultaneously. Since mutation of one chromosome has nothing to do with other chromosomes in this paper, there are no task dependencies between any two threads in these phases. For the crossover part however, threads tend to find a peer to exchange information. Since threads are working on this part together, it is not possible for them to work on original chromosomes directly. Copies have to be made before crossover phase starts. However, these copies still cannot be changed directly because it is possible that two chromosomes choose the same target to communicate with. Since this operation not only reads, but also changes data, working directly on the copies is still not allowed. Therefore, each thread should make another temporary copy for the target chromosome to work on. After crossover phase, the copy for previous group will be used as the group of the last generation. After mutation phase, selection phase needs to compare current chromosome and previous version and decide which one is better, and therefore stays. This process can be directly parallelized since no communication and task dependency exist among threads. Updating the best chromosome needs to search for the minimum one in the adaptive value array for the new chromosome group. This can be implemented in complexity of O(logn) using n processors instead of O(n) done sequentially. The existence of shared memory and cache can reduce this time to an insignificant level, even doing it sequentially in one thread.
3.6 Synchronization in CUDA Based on the task dependency analysis, necessary synchronization points have been detected and inserted as in Fig. 6. Synchronization needs to be addressed at four positions: 1. The copy phase should wait till best value is updated. 2. The crossover phase should wait till copy phase finishes. 3. Updating the best chromosome should wait until the selection phase finishes. 4. On the CPU side, the programmer should place a CUDA synchronization call to wait till all threads are idle, and then output the result. If this is not done, the result will be wrong since CPU does not know what is going inside the GPU. Also, from crossover phase to selection phase, synchronization is not necessary due to the introduction of chromosome copy and algorithm design.
248
S. Chen et al.
Initialize
Initialize
Initialize
Initialize
Copy
Copy
Copy
Copy
Synchronize
Crossover
Crossover
Crossover
Crossover
Mutate
Mutate
Mutate
Mutate
Select
Select
Select
Select
Synchronize
Update Best Value
Synchronize
Reach Condition?
NO
YES Output Results
Fig. 6 Task dependency and synchronization points in CUDA programs
4 Experimental Results and Discussion Both sequential and parallel programs were tested on a machine with two Intel Xeon E5504 Quad-Core CPUs (2.00GHz, 4MB cache) and two NVIDIA Tesla 20-Series C2050 GPUs. Tests have been carefully made to determine how many chromosomes should be generated as a group and how large the termination generations should be. It turns out that we can get better solutions by setting 200 as the chromosome number and 1000 as the termination generation. Values that are larger than these two numbers do not provide further significant improvment to our solutions but increase the running time in a linear speed. In general, the test data can be classified into two types: randomized and clustered, as shown in Figs. 7 and 8, respectively. Both of them occur in real life. However, it is easier for people to tell if clustered cities are well routed than random data through their intuitive observations. Even in programs, the work load of clustered data is smaller than randomized data. When a
CUDA-Based Genetic Algorithm on Traveling Salesman Problem
249
80 70 60 50
Y
40 30 20 10 0 0
10
20
30
40
50
60
70
X
Fig. 7 An example of randomized test data for Genetic Algorithm
90 80 70 60
Y
50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
X
Fig. 8 An example of clustered test data for Genetic Algorithm
good solution is found, it is harder to find a better one for clustered data than for randomized data since only few tiny specific changes can update present the best solution for a cluster, while many more possible changes exist for randomized data. This inherent property associated with these two types of data makes their potential work load different in this paper. When dealing with clustered data, the algorithm will find it hard to update best value after it approaches some sort of line. Hence, the program ends early. On the other hand, the best value tends to update more times for randomized data, which causes a longer average running time. Comparison results are illustrated in Fig. 9. Both GPU and CPU give positive results for the above hypothesis, that is, algorithm on clustered data terminates earlier than that on randomized data. Another fact is that, for both types of data, GPU beats CPU.
250
S. Chen et al.
300 GPU random
RUNNING TIME (Average)
250 CPU random 200
GPU cluster CPU cluster
150
100
50
0 0
100
200
300
400
POPULATION SIZE
Fig. 9 Performance comparison with randomized and clustered data for GA
Questions may be raised about why GA in this paper only gets such insignificant speed up on GPU. As mentioned before, for the synchronization purpose, we only generate one block to run this program. Under CUDA architecture, threads in one block can only be served by one Streaming Multiprocessor (SM). However, in C2050, each GPU has 16 SMs and shared memory and cache are evenly assigned to each SM, which means we only used about 1/16 computing resource on one GPU, and the performance is still better than using CPU (single processor). In future work, we will try to expand the problem scale and keep the whole GPU or clusters busy, and the speed-up will increase significantly.
5 Related Work Computer simulation of evolution started in 1950s with the work of Nils Aall Barricelli [3][4]. Since 1957, Alex Fraser has published a series of papers on simulation of artificial selection of organisms [7][8]. Based on this work, computer simulation of evolution became more popular in 1960s and 1970s. All essential elements of modern genetic algorithms were included in the book by Fraser and Burnell (1970) [9]. Goldberg (1989) first used genetic algorithm to solve the traveling salesman problem [10]. As a method for solving traveling salesman problems, 2-opt was raised by G. A. Croes (1958) in 1950s [5]. Muhlenbein (1989) brought up the concept of PGA (parallel genetic algorithm) [11], which aimed to implement GA on computer clusters. Ismail (2004) implemented PGA using MPI library [12]. In 2008, NVIDIA released latest CUDA SDK2.0 version, which bestowed CUDA much wider range of applications. Stefano et al. (2009) presented a paper about implementing
CUDA-Based Genetic Algorithm on Traveling Salesman Problem
251
a simple GA with CUDA architecture, where sequential code for same algorithm was taken for comparison [6]. Another paper from Petr Pospicha and Jiri (2009) presented a new PGA and implemented it on CUDA [13]. However, performance comparison in Stefano’s work was not based on the same algorithm. In 2010, NVIDIA developed latest version of its GPU architecture, which is called Fermi, for Tesla M2050 and M2070 [2] and corresponding programming guide under these architectures was released [1].
6 Conclusions and Future Work Compared to Stefano’s work in 2009 [6], this paper presents a more complex but parallelizable Genetic Algorithm (not specifically designed for certain GPU architecture) to solve TSP problem. Corresponding sequential C code for the same algorithm is carefully written for the performance comparison. Experimental results show the CUDA program with new Fermi architecture achieves some performance gains, although not so significant. However, considering the massive random memory accesses brought in by this much more complex algorithm and its relatively shorter execution time, this insignificant acceleration indicates that the current GPU architecture may have great potentials in speeding up the existing simulations of group evolution. More advanced performance tuning techniques such as asynchronous communication and zero copy will be applied for further performance gains in the future.
References 1. Nvidia cuda c programming guide 3.1 (2009) 2. Nvidia fermi tuning guide (2009) 3. Barricelli, N.A.: Esempi numerici di processi di evoluzione. Methodos, 45–68 (1954) 4. Barricelli, N.A.: Symbiogenetic evolution processes realized by artificial methods. Methodos 9, 143–182 (1957) 5. Croes, G.A.: A method for solving travling salesman problems. Operations Res. 6(1), 791–812 (1958) 6. Debattisti, S.: Implementation of a simple genetic algorithm within the cuda architecture. In: The Genetic and Evolutionary Computation Conference (2009) 7. Fraser, A.: Simulation of genetic systems by automatic digital computers. Australian Journal of Biological Science 10, 484–499 (1957) 8. Fraser, A., Burnell, D.: Computer models in genetics. Computers and Security 13, 69–78 (1970) 9. Fraser, A., Burnell, D.: Computer Models in Genetics. McGraw-Hill, New York (1970)
252
S. Chen et al.
10. Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning. Lkuwer Academic Publishers (1989) 11. Muhlenbein, H.: Parallel genetic algorithm, population dynamic and combinational optimization. In: Proc. 3rd, International Conference on Genetic Algorithms (1989) 12. Ismail, M.A.: Parallel genetic algorithms (PGAs): master slave paradigm approach using MPI. E-Tech (2004) 13. Pospichal, P., Jaros, J.: Gpu-based acceleration of the genetic algorithm. In: Genetic and Evolutionary Computation Conference (2009)
Design and Implementation of Sensor Framework for U-Healthcare Services Ha eng -Ko n Kim
Abstract. Ubiquito us senso r netwo rk (USN) is one of the important key technologies for future ubiquitous life. USN nodes will be distributed at any place in the future such as street, in-building, campus, and so on. These USN nodes will play various roles like sensing, gathering, transmitting and receiving information about the surround. So, most of these are implemented as wireless communication system with simple hardware architecture. ZigBee protocol is one of the representative USN systems. So, many manufacturers are developing ZigBee hardware platform and their software protocol. To more efficiently implement and deploy USN, we need to know ZigBee protocols and their characteristics. In this paper, we design and Implement a sensor framework systems related to medical and surveillance that are significantly considered for enhancing human life. These are employed under USN environment to construct multiple health care services in which medical sensors are inter-connected to provide efficient management of them. For this configuration, Zigbee based wireless bio-sensors are established for portable measurement in which PSoC technique is utilized for compact implementation. As well, such Zigbee based embedded sensor equipment is devised for UPnP based sensor framework. Keywords: USN, U-healthcare, Zigbee, UPnp, CBD.
1 Introduction USN utilizes wire-line sensor networks and/or wireless sensor networks (WSNs). WSNs are wire networks consisting of interconnected and spatially distributed autonomous devices using sensors to cooperatively monitor Haeng-Kon Kim Department of Computer Engineering, Catholic University of Daegu, Korea e-mail: hangkon@cu.ac.kr
R. Lee (Ed.): Computer and Information Science 2011, SCI 364, pp. 253–261. c Springer-Verlag Berlin Heidelberg 2011 springerlink.com
254
H.-K. Kim
physical or environmental conditions (e.g., temperature, sound, vibration, pressure, motion or pollutants) at different locations. WSNs were generally implemented as isolated networks. Simple design of applications and services based on isolated sensor networks is made by capture and transmission of collected sensed data to designated application systems. Such isolated simple applications and services have been evolving over the years with network advancement, network and service integration, data processing schemes enhanced by business logics and data mining rules, context awareness schemes, development of hardware and software technologies, etc. These technical developments enable the ability to build an intelligent information infrastructure of sensor networks connected to the existing network infrastructure. This information infrastructure has been called ubiquitous sensor network (USN) opening wide possibilities for applications and services based on sensor networks to various customers such as human consumers, public organizations, enterprises and government. USN applications and services are created via the integration of sensor network applications and services into the network infrastructure. They are applied to everyday life in an invisible way as everything is virtually linked by pervasive networking between USN end-users (including machines and humans) and sensor networks, relayed through intermediate networking entities such as application servers, middleware entities, access network entities, and USN gateways. USN applications and services can be used in many civilian application areas such as industrial automation,
Fig. 1 USN Network
Design and Implementation of Sensor Framework
255
home automation, agricultural monitoring, healthcare, environment, pollution and disaster surveillance, homeland security or military field. Many industries invest cost and time for develop ubiquitous computing technology in IT fields. Ubiquitous computing is meant that there are multiple computers are embedded inside human and nature environment and they are inter-connected to be computed for alternative environment as in figure 1.
2 Related Works 2.1 Wireless Sensor Networks Before looking at how wireless sensor networks can be used to assist firefighters in the performance of their duties, it is first necessary to know something about wireless sensor networks in terms of how they work; their capabilities and limitations. A Wireless Sensor Network (WSN) is a network comprised of numerous small independent sensor nodes or motes. They merge a broad range of information technology; hardware, software, networking, and programming methodologies. Wireless Sensor Networks can be applied to a range of applications [1] monitoring of space which includes environmental and habitat monitoring, indoor climate control, surveillance etc.; monitoring things for example structural monitoring, condition-based equipment maintenance etc.; and monitoring the interactions of things with each other and the surrounding space e.g., emergency response, disaster management, healthcare etc. The majority of these applications may be split into two classifications: data collection and event detection. Each mote in a wireless sensor network is a self-contained unit comprised of a power supply (generally batteries), a communication device (radio transceivers), a sensor or sensors, analog-to-digital converters (ADCs), a microprocessor, and data storage [2,3]. The motes self organize themselves, into wireless networks as in figure 2 and data from the motes is relayed to neighboring motes until it reaches the desired destination for processing. Each mote has very limited resources in terms of processing speed, storage capacity and communication bandwidth. In addition, their lifetime is determined by their ability to conserve power. These limitations are a significant factor and must be addressed when designing and implementing a wireless sensor network for a specific application.
2.2 UPnP UPnP which is extensive from Plug-and-Play (PnP) based standard internet protocol popularly includes intelligent electronics, wireless machines and all personal computers to connect Peer-to-Peer in network points of view. Moreover, home network or SOHO and public regions are connected through the internet which provides flexible usage by employing TCP/IP network technology. This can be extensively established to provide PnP functions in
256
H.-K. Kim
Fig. 2 Example of a Flat Network
printers, internet gateways, and home electronics. The devices transfer their ability to active networks through the UPnP services. That is, it uses Universal Control Point to control home applications after detecting and searching related devices. Sensor framework includes sensor searching, registration and deletion, monitoring control functions. This paper uses sensor framework implemented with equal framework and components to able to delete and add in Plug-in structures [3].
3 The Proposed Network Topology 3.1 Concepts Generally, UPnP sensor based framework is a kind of software modules, which is unloaded with UPnP to present UPnP devices for interconnecting. This sensor framework can control the present UPnP devices with the protocol and unloaded devices to translate each protocol. Namely, it is likely to be an emulator of UPnP devices although non-UPnP in reality. We propose the system architectures to be design and implement as shown in figure 3.
Design and Implementation of Sensor Framework
257
Fig. 3 Structure of our systems
The UPnP sensor framework is a device to connect bio-sensor and environmental sensor modules to Zigbee network which cannot UPnP stack. This is able to recognize several sensor modules through UPnP middle-ware, which includes bio and environment modules. Such framework is available to activate different UPnP devices and user control points based on DHCP servers. The UPnP must be constructed with the TCP/IP based UPnP stack to provide connectivity according to utility, flexibility, and standard through UPnP middle-ware. However, the sensor module is not connected with nonIP devices. Thus, we construct the UPnP sensor device modules inside the framework to recognize a virtual UPnP device. Figure 4 illustrates a software
Fig. 4 Structure of Software Module
258
H.-K. Kim
module of the UPnP framework proposed in this paper. The UPnP framework device module: The UPnP standard device model to connect the framework proposed in this paper to the UPnP device. Sensor data processing module: Data to be transferred to the framework is acquired and give its status continuously to the bio-data signal to user application. UPnP sensor equipment module: Based on sensor equipment information to be transferred from its data processing module, the loaded UPnP sensor is practically connected. The derived UPnP sensor device is linked with realistic sensor devices and UPnP control point from the virtual framework.
3.2 Design and Implementation of the UPnP Sensor Network Realization of the UPnP based sensor network to equip Zigbee based senor systems is as follow in figure 5. The framework involves to connect the ports and activate modules of the whole software to acquire data from the sensor systems by the command BridgeStart(). The sensor devices continuously send 66 byte data sets including user ID, sensor type, and bio-data. The framework adds sensor devices listed in Device-ArrayList after acquiring data through the function GetsensorData(). The function SetsensorDevice() equips the sensor devices identified to connect based on Device-ArrayList. The function SetDataXML transforms the FLEX web application into the XML data type in order for the chart presentation. Through this procedure, the sensor equipment is connected with the UPnP control point for sensor device management and control of the user applications which can be identified from the transferred bio-data.
Fig. 5 Software module in our Frameworks
Design and Implementation of Sensor Framework
259
Fig. 6 testing environment for the utilized sensor modules
The UPnP device generally supports the Plug and Play connection to the hardware with its libraries which is possible to be activated under the window based PC or UPnP middle-ware. Figure 6 shows a testing environment for the utilized sensor modules and the web camera inside the home gateway network topology. Figure 7 illustrates the UPnP control point program for identifying and control UPnP network connection. This program is installed inside the home gateway and equips the UPnP to show the web camera. As well, to control
Fig. 7 UPnP Control Point Program
260
H.-K. Kim
the sensor module, the command SetPower can control the power of the sensor systems via the control point action panel. The sensor command is constructed for power on/off action which is available from a Sleep mode of PSoC technique.
Fig. 8 Power control mode
Fig. 9 UI Application
Design and Implementation of Sensor Framework
261
Fig. 8 shows an interfacing display to monitor data from the sensor modules implemented from the FLEX data service 2.0. An established user interface identifies what kind of sensor module is connected including selection and status of users and environment. Figure 9 show the UI applications for the frameworks.
4 Conclusions This paper presents the logical UPnP single network construction which is no limit to connect different application systems provided possibly from standard connectivity and management under embedded USN environments. Main advantages of the proposed system include provision of the standardized connectivity under Zigbee based wireless communication network and effectiveness of device management and control through the UPnP control point program. These proposed topologies are able to extend and change multiple different application systems each other. In future work, we expand this investigation for more rapid and higher service provision in inter-connection of the single network configuration.
References 1. Schwiebert, L., Gupta, S., Weinmann, J.: Research Challenges in Wireless Networks of Biomedical Sensors. In: Proceedings of the 7th Annual International Conference on Mobile Computing and Networking 2. Linnyer Beatrys Ruiz, J.M.S.N., Loureiro, A.A.F.: MANNA: A Management Architecture for Wireless Sensor Networks. IEEE Communications Magazine 41(2), 116–125 (2003) 3. Song, H., Kim, D., Lee, K., Sung, J.: UPnP-Based Sensor Network Management Architecture Real-time and Embedded Systems. Lab Information and Communications University (2007) 4. Eidsvik, A.K., Karlsen, R., Blair, G., Grace, P.: Interfacing remote transaction services using UPnPq. Journal of Computer and System Sciences 74, 158–169 (2008)
Author Index
Abbaspour, Maghsoud Abdallah, Hanˆene Ben Ao, Shan 201 Bouassida, Nadia Chen, Chen, Chen, Chen, Chen,
143 17
111
Matsuo, Tokuro 179 Mei, Jia 1 Miao, Huaikou 1 Min, Ni 191 Motoki, Yosuke 179
17
Hong 47 Huaping 73, 85 Shengbo 1 Su 241 Xuhui 169
Nagpal, Amandeep 127 Novobilski, Andy 241
Dai, Jianhua 201 Davis, Spencer 241 Du, Wencai 213, 227 Fourati, Rahma
Lima, Ricardo M.F. Lina, Qi 191 Liu, Dapeng 59
Oliveira, C´esar A.L.
111
Rafigh, Majid 143 Rahman, Rashedur M.
17
Hu, Dewen 169 Hu, Gongzhu 155
Sabat, Cec´ılia L. 111 Saito, Yoshihito 179 Shen, Hui 169 Silva, Nat´ alia C. 111 Singh, Balraj 127
Imam, Toukir
Takahashi, Satoshi
Gao, Honghao 1 Gao, Lijin 35
95
Jiang, Hai 241 Johal, Hartinder Singh
Wang, Yunqiong
179 35
127
Ke, Ming 169 Kim, Haeng-Kon 253 Krishan, Kewal 127 Lee, Matthew K.O. Lee, Roger 155
95
73, 85
Xin, Qinling 213, 227 Xu, Shaochun 59, 213, 227 Xu, Tianwei 35 Yang, Rongfang 35 Yongqiu, Xie 191 Yuan, Jinhui 47
264 Yu, Hao 191 Yunpeng, Li 191 Zhang, Jin 155 Zhang, Kem Z.K. 73, 85 Zhao, Sesia J. 73, 85
Author Index Zhou, Hongwei 47 Zhou, Hui 213, 227 Zhou, Juxiang 35 Zhou, Xiaolin 169 Zhou, Zongtan 169 Zhu, Ge 201